@noggn/mcp

Local stdio MCP server for the Checklist API, so an agent can act on the API for a single organization. Generated at startup from the bundled OpenAPI spec and tiered for token efficiency: the run-time hot path (assignment loop + checklist/template reads + create/import) stays as ~12 typed tools, and every other operation is reachable through one generic checklist_request tool — full API parity retained (all 54 ops callable) at ~⅓ the every-session token cost. The checklist://openapi resource serves exact schemas on demand. The org's secret key (CHECKLIST_API_KEY) scopes every action to that account; calls dispatch through @noggn/checklists, inheriting auth, version pinning, auto-idempotency, and retries.

Run it

# published:
CHECKLIST_API_KEY=sk_… npx @noggn/mcp

# workspace dev (against local API):
CHECKLIST_API_KEY=sk_… CHECKLIST_BASE_URL=http://127.0.0.1:8787 pnpm start

Environment

| Var | Required | Notes | |---|---|---| | CHECKLIST_API_KEY | yes | the sk_… secret; scopes all actions to that account | | CHECKLIST_BASE_URL | no | API base URL (defaults to local dev) | | CHECKLIST_VERSION | no | pin the Checklist-Version | | CHECKLIST_OPENAPI | no | override the bundled spec path |

The spec is bundled into dist/openapi.json at build; dev reads the sibling packages/api/openapi.json.

MCP client config (Claude Desktop / Cursor)

{
  "mcpServers": {
    "checklist": {
      "command": "npx",
      "args": ["@noggn/mcp"],
      "env": { "CHECKLIST_API_KEY": "sk_…" }
    }
  }
}

Programmatic (AI SDK v6 MCP client):

import { experimental_createMCPClient as createMCPClient } from 'ai'

const mcp = await createMCPClient({
  transport: {
    type: 'stdio',
    command: 'npx',
    args: ['@noggn/mcp'],
    env: { CHECKLIST_API_KEY: process.env.CHECKLIST_API_KEY! },
  },
})
const tools = await mcp.tools()

See AGENTS.md for how tools are generated and the transport-agnostic server design.